Replace composite dispatch with `CompositeExplicitAutograd` (#64641)
authorYukio Siraichi <yukio.siraichi@gmail.com>
Tue, 14 Sep 2021 14:55:13 +0000 (07:55 -0700)
committerFacebook GitHub Bot <facebook-github-bot@users.noreply.github.com>
Tue, 14 Sep 2021 14:56:54 +0000 (07:56 -0700)
commit2853c7da2243611c6fbd8b78b93c0cd34f0f1bc1
treed1d6bedc3b2021fbf99fad09c447ed2bd6b56c07
parent09d221e8d439bc748b162c028f7eece202688adf
Replace composite dispatch with `CompositeExplicitAutograd` (#64641)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/64641

`sum`, `mean`, and `norm` were ported to structured kernels in #61642, #61643, and #62711,
respectively. Those PRs changed related overlads into composite kernels. However, their
dispatch section remained the same, when they really should be marked as
`CompositeExplicitAutograd`. This PR fixes this issue.

Test Plan: Imported from OSS

Reviewed By: ngimel

Differential Revision: D30867122

Pulled By: ezyang

fbshipit-source-id: b951aee41a3cab9ca546df826a285d60013e3b3a
aten/src/ATen/native/ReduceOps.cpp
aten/src/ATen/native/native_functions.yaml
aten/src/ATen/native/quantized/cpu/qreduction.cpp